Embodiments include a method for pruning anchor points from a feature map generated from a Light Detections And Ranging (LiDAR) point cloud, the method comprising: receiving, by a navigation system, a LiDAR point cloud from a LiDAR sensor, the LiDAR point cloud comprising data representing one or more objects in physical surroundings detected by the LiDAR sensor; extracting, by the navigation system, a feature map from the LiDAR point cloud, the feature map comprising a plurality of anchor points, each anchor point defined by an anchor box; smoothing, by the navigation system, the extracted feature map; determining, by the navigation system, density of pixels within the anchor box of each anchor point; and pruning, by the navigation system, anchor points from the feature map based on a plurality of factors related to the determined density of pixels within the box of each anchor point.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method for pruning anchor points from a feature map generated from a Light Detections And Ranging (LiDAR) point cloud, the method comprising: receiving, by a navigation system, a LiDAR point cloud from a LiDAR sensor, the LiDAR point cloud comprising data representing one or more objects in physical surroundings detected by the LiDAR sensor; extracting, by the navigation system, a feature map from the LiDAR point cloud, the feature map comprising a plurality of anchor points, each anchor point defined by an anchor box; smoothing, by the navigation system, the extracted feature map; determining, by the navigation system, a density of pixels within the anchor box of each anchor point; and pruning, by the navigation system, anchor points from the feature map based on a plurality of factors related to the determined density of pixels within the anchor box of each anchor point.
This invention relates to improving object detection in autonomous navigation systems using Light Detection and Ranging (LiDAR) sensors. The problem addressed is the computational inefficiency and noise in feature maps derived from LiDAR point clouds, which can lead to inaccurate object detection and navigation decisions. The solution involves a method for pruning anchor points from a feature map generated from a LiDAR point cloud to enhance detection accuracy and reduce processing overhead. The method begins with receiving a LiDAR point cloud from a LiDAR sensor, which contains data representing objects in the physical surroundings. A feature map is then extracted from the point cloud, consisting of multiple anchor points, each defined by an anchor box. The feature map is smoothed to reduce noise. The system then determines the density of pixels within the anchor box of each anchor point. Finally, anchor points are pruned from the feature map based on factors related to the determined pixel density, such as sparsity or clustering, to retain only the most relevant anchor points for object detection. This process optimizes the feature map by removing redundant or low-density anchor points, improving the efficiency and accuracy of object detection in autonomous navigation systems.
2. The method of claim 1 , wherein smoothing the extracted feature map comprises applying a Gaussian smoothing to the extracted feature map.
A method for processing feature maps in image analysis involves smoothing the extracted feature map to reduce noise and enhance feature detection. The method applies Gaussian smoothing to the extracted feature map, which involves convolving the feature map with a Gaussian kernel. This process blurs the feature map, attenuating high-frequency noise while preserving the overall structure of the features. Gaussian smoothing is particularly useful in computer vision and machine learning applications where feature maps are derived from convolutional neural networks or other image processing techniques. The smoothing step helps improve the robustness of subsequent analysis, such as object detection, segmentation, or classification, by reducing the impact of spurious variations in the feature map. The Gaussian kernel's standard deviation and size can be adjusted to control the degree of smoothing, allowing for fine-tuning based on the specific requirements of the application. This technique is commonly used in preprocessing stages to enhance the quality of feature maps before further processing or analysis.
3. The method of claim 1 , wherein determining the density of pixels within the anchor box of each anchor point comprises computing a sum of pixels within the anchor box of each anchor point.
This invention relates to image processing, specifically to methods for analyzing pixel density within predefined regions of an image. The problem addressed is efficiently determining the density of pixels within specific anchor boxes to improve object detection or segmentation tasks in computer vision systems. The method involves computing the sum of pixel values within an anchor box associated with each anchor point in an image. Anchor points are predefined reference locations used to define regions of interest, and anchor boxes are rectangular or similarly shaped areas centered around these points. By summing the pixel values within each anchor box, the method quantifies the density of pixels, which can indicate the presence or absence of objects, edges, or other features. This approach is useful in applications such as object detection, where pixel density within anchor boxes helps identify potential objects. The method may be part of a larger system that uses these density values to refine anchor box positions, adjust detection algorithms, or improve segmentation accuracy. The invention focuses on optimizing computational efficiency by simplifying the density calculation to a summation operation, reducing complexity while maintaining accuracy. This technique is applicable in real-time image processing, autonomous systems, and other fields requiring rapid analysis of visual data.
4. The method of claim 3 , wherein determining the density of pixels within the anchor box of each anchor point further comprises computing an average pixel value for the anchor box of each anchor point.
This invention relates to image processing, specifically to methods for analyzing pixel density within predefined regions of an image. The problem addressed is the need for efficient and accurate computation of pixel density in anchor boxes, which are predefined regions used in image analysis tasks such as object detection or feature extraction. Traditional methods may struggle with computational efficiency or accuracy when determining pixel density, particularly in large or complex images. The invention improves upon prior art by computing an average pixel value for each anchor box. An anchor point is a reference location within an image, and an anchor box is a predefined region around that point. The method involves calculating the average pixel value of all pixels within each anchor box, which provides a more robust and computationally efficient measure of pixel density compared to other techniques. This approach helps in accurately identifying regions of interest, such as objects or features, by leveraging the average pixel value as a density metric. The method can be applied in various image processing applications, including computer vision, machine learning, and automated image analysis, where precise density calculations are critical for performance. By focusing on the average pixel value, the invention ensures consistency and reliability in density measurements, improving the overall accuracy of image analysis tasks.
5. The method of claim 4 , wherein determining the density of pixels within the anchor box of each anchor point further comprises computing a center pixel value for the anchor box of each anchor point.
This invention relates to image processing, specifically to methods for analyzing pixel density within defined regions of an image to improve object detection or segmentation. The problem addressed is the need for efficient and accurate computation of pixel density within anchor boxes, which are predefined regions used to identify objects in images. Traditional methods may struggle with computational efficiency or accuracy when determining pixel density, particularly in complex scenes with varying object sizes and shapes. The method involves computing a center pixel value for each anchor box associated with an anchor point. An anchor point is a reference location in the image where an anchor box is defined, typically used in object detection frameworks like Faster R-CNN or YOLO. The anchor box is a rectangular region centered around the anchor point, and its size may vary depending on the application. By calculating the center pixel value, the method provides a simplified yet effective way to assess pixel density within the anchor box, which can be used to refine object detection or segmentation tasks. This approach reduces computational overhead compared to analyzing all pixels within the box while maintaining accuracy in density estimation. The method is particularly useful in real-time applications where processing speed is critical, such as autonomous driving or surveillance systems.
6. The method of claim 5 , wherein pruning anchor points from the feature map based on the plurality of factors related to the determined density of pixels within the anchor box of each anchor point comprises pruning anchor points having an anchor box with an average pixel value below a predetermined threshold for average pixel value.
This invention relates to image processing, specifically to improving object detection by optimizing anchor points in a feature map. The problem addressed is the computational inefficiency and accuracy limitations in object detection systems caused by redundant or low-quality anchor points, which are predefined reference boxes used to detect objects in images. The method involves pruning anchor points from a feature map based on factors related to the density of pixels within the anchor box of each anchor point. Specifically, anchor points are removed if their anchor box has an average pixel value below a predetermined threshold. This ensures that only anchor points with sufficient pixel density, and thus higher likelihood of containing meaningful features, are retained. The pruning process helps reduce computational overhead and improves detection accuracy by focusing on relevant regions of the image. Additionally, the method may include generating a feature map from an input image, determining the density of pixels within the anchor box of each anchor point, and selecting anchor points based on the determined density. The pruning step further refines the selection by eliminating anchor points with insufficient pixel density, thereby optimizing the detection process. This approach enhances the efficiency and effectiveness of object detection in image processing systems.
7. The method of claim 6 , wherein pruning anchor points from the feature map based on the plurality of factors related to the determined density of pixels within the anchor box of each anchor point further comprises pruning anchor points having an anchor box with a center pixel value below a predetermined threshold for center pixel value.
This invention relates to image processing, specifically to methods for optimizing object detection by pruning anchor points in a feature map. The problem addressed is the computational inefficiency and redundancy in object detection systems, where numerous anchor points are generated but many are irrelevant or low-quality, leading to unnecessary processing. The method involves analyzing a feature map to identify anchor points, each associated with an anchor box representing a potential object location. To improve efficiency, anchor points are pruned based on multiple factors, including the density of pixels within each anchor box. A key aspect is evaluating the center pixel value of each anchor box. If the center pixel value falls below a predetermined threshold, the corresponding anchor point is removed from consideration. This step ensures that only anchor points with sufficient pixel density and relevance are retained, reducing computational overhead and improving detection accuracy. The pruning process is part of a broader method for object detection that involves generating anchor points, evaluating their quality, and selecting the most promising candidates for further processing. By focusing on anchor points with high-quality center pixels, the system avoids processing low-value candidates, enhancing both speed and accuracy in object detection tasks. This approach is particularly useful in applications requiring real-time processing, such as autonomous vehicles or surveillance systems.
8. A navigation system comprising: a processor; and a memory coupled with and readable by the memory and storing therein a set of instructions which, when executed by the processor, causes the processor to prune anchor points from a feature map generated from a Light Detections And Ranging (LiDAR) point cloud by: receiving, by a navigation system, a LiDAR point cloud from a LiDAR sensor, the LiDAR point cloud comprising data representing one or more objects in physical surroundings detected by the LiDAR sensor; extracting, by the navigation system, a feature map from the LiDAR point cloud, the feature map comprising a plurality of anchor points, each anchor point defined by an anchor box; smoothing, by the navigation system, the extracted feature map; determining, by the navigation system, a density of pixels within the anchor box of each anchor point; and pruning, by the navigation system, anchor points from the feature map based on a plurality of factors related to the determined density of pixels within the anchor box of each anchor point.
This invention relates to navigation systems that process Light Detection and Ranging (LiDAR) point clouds to improve object detection and mapping. The system addresses challenges in accurately identifying and filtering relevant features from LiDAR data, which is critical for autonomous navigation and obstacle avoidance. The system includes a processor and memory storing instructions that, when executed, perform the following steps. A LiDAR sensor captures a point cloud representing objects in the physical surroundings. The system extracts a feature map from this point cloud, where the map consists of multiple anchor points, each defined by an anchor box. The feature map is then smoothed to reduce noise. The system determines the pixel density within each anchor box and prunes anchor points based on this density and other related factors. This pruning step removes irrelevant or redundant features, enhancing the accuracy and efficiency of object detection. By dynamically adjusting the feature map, the system improves navigation performance in environments with complex or cluttered surroundings. The method ensures that only meaningful features are retained, reducing computational overhead and improving real-time decision-making for autonomous systems.
9. The navigation system of claim 8 , wherein smoothing the extracted feature map comprises applying a Gaussian smoothing to the extracted feature map.
The invention relates to navigation systems that process feature maps for improved navigation accuracy. The problem addressed is the presence of noise or irregularities in extracted feature maps, which can degrade the performance of navigation algorithms. To solve this, the system applies Gaussian smoothing to the extracted feature map. Gaussian smoothing is a mathematical technique that reduces noise by averaging pixel values in the feature map using a Gaussian-weighted kernel. This process enhances the clarity and reliability of the feature map, making it more suitable for navigation tasks such as localization, path planning, or obstacle detection. The smoothed feature map can then be used by the navigation system to make more accurate decisions, such as determining the position of a vehicle or robot within an environment or identifying optimal paths while avoiding obstacles. The Gaussian smoothing step ensures that the feature map is free from high-frequency noise, which could otherwise lead to incorrect navigation outputs. This technique is particularly useful in environments where sensor data is prone to noise, such as in autonomous driving or robotic navigation systems.
10. The navigation system of claim 8 , wherein determining the density of pixels within the anchor box of each anchor point comprises computing a sum of pixels within the anchor box of each anchor point.
The invention relates to navigation systems that use visual data to determine the density of pixels within predefined regions, known as anchor boxes, to improve navigation accuracy. The system addresses the challenge of accurately identifying and tracking visual features in dynamic environments, which is critical for autonomous navigation and robotics. By analyzing pixel density within anchor boxes, the system can enhance the reliability of feature detection and matching, reducing errors caused by occlusions, lighting changes, or other environmental variations. The navigation system includes a method for processing visual data captured by sensors, such as cameras, to identify anchor points in the environment. Each anchor point is associated with an anchor box, a predefined region of pixels around the point. The system computes the density of pixels within each anchor box by summing the pixel values, which provides a quantitative measure of the visual information present in that region. This density calculation helps distinguish between areas with high feature density (e.g., textured surfaces) and low feature density (e.g., uniform backgrounds), enabling the system to prioritize regions with more reliable visual features for navigation. The system may also include additional techniques for refining anchor point selection, such as filtering out low-density regions or adjusting anchor box sizes dynamically. By integrating pixel density analysis with other navigation algorithms, the system improves the robustness and accuracy of visual-based navigation in real-world applications.
11. The navigation system of claim 10 , wherein determining the density of pixels within the anchor box of each anchor point further comprises computing an average pixel value for the anchor box of each anchor point.
The invention relates to navigation systems that use visual data, such as images or video, to determine the position or orientation of a device. A key challenge in such systems is accurately identifying and tracking anchor points within the visual data to ensure reliable navigation. The invention addresses this by improving the way anchor points are processed to enhance navigation accuracy. The system identifies anchor points within a visual frame and defines an anchor box around each point. To refine the anchor point data, the system computes the density of pixels within each anchor box. This involves calculating an average pixel value for the pixels inside the anchor box. The average pixel value helps determine the significance or reliability of each anchor point, allowing the system to filter out less reliable points and improve navigation accuracy. The system may also adjust the anchor box size or position based on the computed pixel density to better capture relevant visual features. By analyzing pixel density and average pixel values within anchor boxes, the system enhances the robustness of visual navigation, reducing errors caused by noise or irrelevant features in the visual data. This method ensures that only the most relevant anchor points are used for navigation calculations, improving overall system performance.
12. The navigation system of claim 11 , wherein determining the density of pixels within the anchor box of each anchor point further comprises computing a center pixel value for the anchor box of each anchor point.
The invention relates to navigation systems that use visual data, such as images or video, to determine the position and orientation of a device. A key challenge in such systems is accurately identifying and tracking anchor points in the environment to enable precise navigation. The system addresses this by analyzing the density of pixels within predefined anchor boxes around each anchor point. To improve accuracy, the system computes a center pixel value for each anchor box. This center pixel value helps refine the density calculation, ensuring that the system can reliably detect and track anchor points even in complex or cluttered environments. The method involves capturing visual data, identifying potential anchor points, and processing the pixel data within their respective anchor boxes to derive the center pixel value. This refined density measurement enhances the system's ability to maintain accurate navigation in dynamic or visually challenging conditions. The approach is particularly useful in applications like autonomous vehicles, robotics, and augmented reality, where precise spatial awareness is critical. By focusing on pixel density and center pixel values, the system improves robustness against noise and occlusions, ensuring reliable navigation performance.
13. The navigation system of claim 12 , wherein pruning anchor points from the feature map based on the plurality of factors related to the determined density of pixels within the anchor box of each anchor point comprises pruning anchor points having an anchor box with an average pixel value below a predetermined threshold for average pixel value.
This invention relates to navigation systems that use feature maps to improve localization accuracy. The problem addressed is the computational inefficiency and noise in feature maps caused by excessive or irrelevant anchor points, which can degrade navigation performance. The system includes a feature map with anchor points, each associated with an anchor box containing pixels. The system determines the density of pixels within each anchor box and prunes anchor points based on multiple factors, including the average pixel value within the anchor box. Specifically, anchor points with an average pixel value below a predetermined threshold are removed to reduce noise and improve processing efficiency. The system may also adjust the size of anchor boxes or modify anchor point positions to enhance feature map accuracy. The pruning process ensures that only relevant and high-quality anchor points remain, improving the reliability of navigation calculations. This approach optimizes computational resources while maintaining or improving localization precision.
14. The navigation system of claim 13 , wherein pruning anchor points from the feature map based on the plurality of factors related to the determined density of pixels within the anchor box of each anchor point further comprises pruning anchor points having an anchor box with a center pixel value below a predetermined threshold for center pixel value.
A navigation system for autonomous vehicles or robotic systems processes sensor data to generate a feature map with anchor points representing potential navigation landmarks. The system evaluates the density of pixels within an anchor box around each anchor point to determine relevance. To optimize processing, the system prunes anchor points based on multiple factors, including the density of pixels within their anchor boxes. Specifically, anchor points are removed if the center pixel value of their anchor box falls below a predetermined threshold, ensuring only meaningful landmarks are retained for navigation. This selective pruning reduces computational load while maintaining accuracy in path planning and obstacle avoidance. The system dynamically adjusts the feature map by filtering out low-relevance anchor points, improving efficiency in real-time navigation tasks. The threshold for center pixel value acts as a quality gate, ensuring only sufficiently distinct features are considered, which enhances the reliability of the navigation system in varying environments. This approach balances performance and precision, making it suitable for applications requiring robust spatial awareness.
15. A vehicle comprising: a Light Detection And Ranging (LiDAR) sensor; a navigation system coupled with the LiDAR sensor and comprising a processor and a memory coupled with and readable by the processor and storing therein a set of instructions which, when executed by the processor, causes the processor to prune anchor points from a feature map generated from a Light Detections And Ranging (LiDAR) point cloud by: receiving, by a navigation system, a LiDAR point cloud from a LiDAR sensor, the LiDAR point cloud comprising data representing one or more objects in physical surroundings detected by the LiDAR sensor; extracting, by the navigation system, a feature map from the LiDAR point cloud, the feature map comprising a plurality of anchor points, each anchor point defined by an anchor box; smoothing, by the navigation system, the extracted feature map; determining, by the navigation system, a density of pixels within the anchor box of each anchor point; and pruning, by the navigation system, anchor points from the feature map based on a plurality of factors related to the determined density of pixels within the anchor box of each anchor point.
This invention relates to autonomous vehicle navigation systems that use Light Detection and Ranging (LiDAR) sensors to improve object detection and mapping. The system addresses the challenge of processing large volumes of LiDAR point cloud data to generate accurate and efficient feature maps for navigation. A LiDAR sensor captures data representing objects in the vehicle's surroundings, which is then processed by a navigation system. The navigation system extracts a feature map from the LiDAR point cloud, where the map consists of multiple anchor points, each defined by an anchor box. The system smooths the feature map to reduce noise and then determines the pixel density within each anchor box. Based on this density and other factors, the system prunes or removes anchor points from the feature map. This pruning process optimizes the feature map by eliminating redundant or low-density anchor points, improving computational efficiency and accuracy in object detection and navigation. The navigation system includes a processor and memory storing instructions to execute these steps, ensuring real-time processing of LiDAR data for autonomous driving applications.
16. The vehicle of claim 15 , wherein smoothing the extracted feature map comprises applying a Gaussian smoothing to the extracted feature map.
This invention relates to vehicle systems that process sensor data, particularly for object detection or scene understanding. The problem addressed is improving the accuracy and reliability of feature extraction from sensor data, such as images or LiDAR point clouds, by reducing noise and enhancing relevant features before further processing. The system includes a sensor, such as a camera or LiDAR, that captures environmental data. A feature extraction module processes this data to generate a feature map, which highlights key characteristics like edges, textures, or object boundaries. To improve the feature map, a smoothing module applies Gaussian smoothing, which reduces high-frequency noise while preserving important structural information. This smoothed feature map is then used by a downstream processing module, such as an object detection or classification algorithm, to make more accurate decisions. Gaussian smoothing is a mathematical technique that convolves the feature map with a Gaussian kernel, effectively blurring the data to suppress noise while maintaining the integrity of larger, more meaningful features. This step enhances the robustness of subsequent analysis, particularly in challenging conditions like low light or adverse weather. The system may also include additional preprocessing steps, such as normalization or contrast enhancement, to further refine the input data before feature extraction. The overall goal is to improve the reliability of autonomous vehicle perception systems by ensuring high-quality feature representation.
17. The vehicle of claim 15 , wherein determining the density of pixels within the anchor box of each anchor point comprises computing a sum of pixels within the anchor box of each anchor point.
This invention relates to a vehicle equipped with a system for detecting objects in an image using anchor points and anchor boxes. The system addresses the challenge of accurately identifying and localizing objects in images captured by vehicle sensors, such as cameras, to support autonomous driving or advanced driver-assistance systems (ADAS). The vehicle includes an image sensor configured to capture images of a scene and a processor that processes these images to detect objects. The processor generates a plurality of anchor points, each associated with an anchor box representing a potential object location. To refine detection accuracy, the processor determines the density of pixels within each anchor box by computing the sum of pixels in the box. This pixel density calculation helps assess the likelihood of an object being present, improving detection reliability. The system may also adjust the anchor boxes based on the computed density, optimizing object detection performance. The invention enhances object detection in real-time applications, ensuring safer and more efficient vehicle operation.
18. The vehicle of claim 17 , wherein determining the density of pixels within the anchor box of each anchor point further comprises computing an average pixel value for the anchor box of each anchor point.
This invention relates to a vehicle equipped with a system for detecting objects in an environment using image processing. The system addresses the challenge of accurately identifying and localizing objects, such as pedestrians or obstacles, in real-time to enhance safety and autonomous navigation. The vehicle includes a camera or sensor array that captures images of the surrounding environment. The system processes these images by dividing them into a grid of anchor points, each associated with an anchor box—a predefined region of interest. For each anchor point, the system determines the density of pixels within its corresponding anchor box. This density calculation involves computing an average pixel value for the anchor box, which helps assess the likelihood of an object being present. The system then uses this density information to refine object detection, improving accuracy and reducing false positives. The vehicle may further adjust the anchor box size or shape dynamically based on environmental conditions or object characteristics to enhance detection performance. This approach enables reliable object detection in varying lighting and weather conditions, supporting advanced driver-assistance systems and autonomous driving capabilities.
19. The vehicle of claim 18 , wherein determining the density of pixels within the anchor box of each anchor point further comprises computing a center pixel value for the anchor box of each anchor point.
The invention relates to a vehicle equipped with a vision-based object detection system that processes image data to identify and track objects in its surroundings. The system uses anchor points within a defined region to analyze pixel density, enabling more accurate object localization. Specifically, the system computes a center pixel value for each anchor box associated with an anchor point to refine the density calculation. This center pixel value serves as a reference point within the anchor box, improving the precision of object detection by accounting for variations in pixel distribution. The method involves analyzing the anchor box of each anchor point to determine pixel density, where the center pixel value is derived from the central region of the anchor box. This approach enhances the system's ability to distinguish objects from background noise, particularly in dynamic or cluttered environments. The computed center pixel value helps mitigate errors caused by uneven pixel distributions, ensuring more reliable object detection and tracking for autonomous driving or advanced driver-assistance systems.
20. The vehicle of claim 19 , wherein pruning anchor points from the feature map based on the plurality of factors related to the determined density of pixels within the anchor box of each anchor point comprises: pruning anchor points having an anchor box with an average pixel value below a predetermined threshold for average pixel value; and pruning anchor points having an anchor box with a center pixel value below a predetermined threshold for center pixel value.
This invention relates to object detection in autonomous vehicles, specifically improving the efficiency of anchor point selection in feature maps. The problem addressed is the computational overhead and noise in object detection systems caused by excessive or irrelevant anchor points, which can degrade performance and accuracy. The system processes a feature map generated from sensor data, such as images or LiDAR scans, to identify potential object locations. Anchor points are candidate regions within the feature map where objects might be present. To optimize detection, the system prunes anchor points based on pixel density within their anchor boxes. Two key pruning criteria are applied: first, anchor points with an average pixel value below a predefined threshold are removed, as low average values indicate sparse or irrelevant regions. Second, anchor points with a center pixel value below another threshold are discarded, as this suggests the region lacks a strong central feature, which is critical for object detection. By filtering anchor points using these density-based thresholds, the system reduces computational load and improves detection accuracy by focusing on regions with meaningful pixel activity. This approach enhances real-time processing capabilities, which is crucial for autonomous vehicle navigation and safety. The method ensures that only relevant anchor points are retained, improving the efficiency of subsequent object detection and classification stages.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
February 4, 2020
February 22, 2022
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.